Re: [PATCH v2] gdb/amd-dbgapi-target: implement the 'interrupt' target op

[Pedro Alves <pedro@palves.net>]

Hi!

On 2026-05-12 18:33, Tankut Baris Aktemur wrote:

>  
> +void
> +amd_dbgapi_target::interrupt ()
> +{
> +  stop (minus_one_ptid);
> +}

Two issues here:

#1 - Note that this changes the behavior of the CPU target.  

We push the dbgapi target on the stack even before GPU debugging is active.  

Before the patch, "interrupt" stops with SIGINT.

After the patch, "interrupt" stops with SIGINT vs "stopped" depending on whether
dbgapi is built into GDB and active, or not.

Ultimately, I want to make the linux target interrupt with "stopped" too, and my
larger "ctrl-c rework" series does that:
  https://inbox.sourceware.org/gdb-patches/20210603190243.2609886-17-pedro@palves.net/

but maybe we can have a simpler localized patch in linux-nat.c meanwhile, similar to
what I did for Windows recently, here:
  https://inbox.sourceware.org/gdb-patches/20260505122434.1444507-1-pedro@palves.net/


#2 - That is going to iterate over all threads and queue a stop for every thread.  That is
incorrect in all-stop mode.  The assumption is that there is only one single event queued,
and then infrun stops all threads when it handles the event.

E.g., if I try without scheduler-locking, I see:

(gdb) c&
Continuing.
(gdb) info threads
  Id   Target Id                                             Frame 
  1    Thread 0x7ffff62a8540 (LWP 3244715) "interrupt-singl" (running)
  2    Thread 0x7fffe8dff6c0 (LWP 3244828) "interrupt-singl" (running)
  3    Thread 0x7fffe3fff6c0 (LWP 3244829) "interrupt-singl" (running)
  6    Thread 0x7ffff5adf6c0 (LWP 3244832) "interrupt-singl" (running)
* 7    AMDGPU Wave 1:2:1:1 (0,0,0)/0 "kern"                  (running)
(gdb) interrupt 
(gdb) ❌️ wave_stop for wave_1 failed (The wave has an outstanding stop request)
(gdb) ❌️ wave_stop for wave_1 failed (The wave has an outstanding stop request)
(gdb) ❌️ wave_stop for wave_1 failed (The wave has an outstanding stop request)
(gdb) ❌️ wave_stop for wave_1 failed (The wave has an outstanding stop request)
(gdb) 
Thread 7 "kern" stopped.
0x00007ffff7fa7d60 in loop () at /home/pedro/rocm/gdb/build/gdb/testsuite/../../../src/gdb/testsuite/gdb.rocm/interrupt-single.cpp:25
25          __builtin_amdgcn_s_sleep (8);


It's probably also possible to trigger badness in the CPU side.

Your test should definitely cover this, not just schedlock on.

Pedro Alves